Singularity Institute

Cynthia Breazeal discusses how to build a lifelike and seemingly emotional robot.

Singularity Institute: How do you build a machine that can interact with humans?

Cynthia Breazeal: ...Our endeavor is really more like building an artificial creature than building just a machine that is more capable. So we look to science.

We look to primarily developmental psychology but all forms of psychology, neuroscience, ethology to gain ideas and inspiration for how you build a machine that has these, not only life like properties, but can think and make decisions and learn, and again, for us in particular, interact and relate to others to be able to coordinate their behaviors with others. One of the big challenges we have been grappling with over the last few years is how do you give machines a theory of other minds.

SI:So how do you build machines that can actually understand people's behaviors in terms of psychological states?

CB: So not just in terms of activity patterns of what you are doing, but might your intentions be. What might your motives be. What might you beliefs be? We have looked a lot to neuroscience and psychology to try to understand how do people do this, how do children develop the ability to do this, and trying to model that in and emulate that in machines.

So one of the big topics I'll be talking about is perspective taking and theory of mind and how if you have that you can basically much of how we coordinate our minds is through coordinating bodies and vice versa.

How do you actually model that in machine and be able to demonstrate that? But it's a very different question, I think, particularly for robotics because most of robotics is really been about how do you build machines that interact with things? How do you build robots that can pick up cups and assemble structures and navigate Martian landscapes.

And when you think about human society, I mean, there is people everywhere. It's profoundly social and it's a whole new set of questions when you think about how do you build a machine that is capable of interactions with people because people are not only governed by the laws of physics, they are governed by having minds.

And now these machines have to be able to understand people on those terms, as well. So that's really opening up a whole new set of questions.

SI: Can a machine have its own perspective?

CB: By perspective, basically perspective taking is basically sort of the proverbial to understand someone you have to be able to stand in their shoes. So it could be something as straightforward as visual perspective.

If you were able to appreciate a visual scene from another person's perspective, you might be able to infer what they might believe to be true about that scene.

But if you go much further than that. If you could recognize the actions that people are taking within a context from a person's perspective, you might be able to basically use yourself as a simulator to think about if I were doing his actions in that context, what might my goals be?

What might my motives be? And that's a very similar process to how our robots are trying to at least give a first approximation of what a human collaborator's intentions might be, and beliefs be and motives.

SI: Aren't you just programming emotion?

To get these robots to the point where they can really interact with people and these more psychological terms, some of it certainly involves modeling cognitive processes. That could be akin to programming, programming computational models.

But learning and development also plays a profound role. So another side of our work really looks to again infant social development and looking at well, how does that develop. Because, I mean, humans are profoundly social but we are not born that way. It takes us many, many years to develop that competence.

A lot of it is because we acquire social interactions with us in order to acquire that ability. So a lot of what we are doing with robots as well is looking at social forms of learning, such as imitation. And there are a number of prominent development psychologists who feel that one way that infants learn how to simulate others, how to be able to learn this sort of theory of other minds, begins with imitation.

So if you as an infant see other adults around you make the hypothesis, implicit hypothesis that they are like me. I'm like them. They can mimic me. I can try to mimic them. If I can mimic them then somehow I can try to get into their head, so to speak, again this sort of perspective taking ability.

It's intriguing to think about the role of these social forms of learning and developing these more sophisticated ways of thinking and understanding people along the social dimension. Another example is social referencing... This is a form of learning that of course we do as adults, but the first time it really sort of peaks strongly is at one year of age.

And social referencing is basically if I am encountering a novel thing, I don't know if it is good or bad. I don't know what to make of it. Infants often look to their mothers to some sort of trusted care-giving figure and see what their reaction is.

If mom was looking a little frightened, then I might understand that her reaction is about this thing or this context and then I might start to assume or adopt that same appraisal to help guide my own behavior.

And this is, of course, profoundly, I think, important for infants, because there is so much in the world that we don't know if it's good or bad for us. I think similarly for robots, you can imagine trying to bootstrap this sort of intelligence in the world.

They are not suddenly going to know off the bat if things are good or bad or dangerous or safe. They might have to look to people or perhaps other robots to be able to make that appraisal. So I think these forms of social learning, learning from others, learning about others is really fundamental.

SI: Can machine learned behavior be copied and transferred to another machine?

If you talk about a computational entity you definitely also have the opportunity that once one robot learns something that you could potentially pass that on, whether it is through beaming it over wireless communication of what not. But somehow you can pass that on to another robot or another machine to bootstrap its ability to have that skill.

Now, depending on the machine, if it has similar morphology, similar sensing, I mean, it might have to then use that as a starting point and then learn and experience and explore, in order to hone or tune that. So I think certainly within computational entities, that is certainly possible.

But I think for me personally in the bigger picture, there are things that I think or that I would like these machines to know other than tasks and facts and so forth, but human values.

What's considered appropriate and inappropriate. I think these sort of culturally, socially situated and contextualized bits of knowledge through interaction I think are also important.

I guess people are always talking about you go to the Internet and pick up that stuff, but one of the things I'm trying to do with these robots in particular, as opposed to doing it through the abstract process of symbol manipulation.

These robots are actually being designed to try to experience what other people are experiencing, so it's a much more empathetic way of understanding others.

And, I think, people, certainly do both and I think they both have a role, but I think, again, if you're thinking about intelligent machines trying to relate to people, if they can empathize with us, I have a hunch that they're probably going to be more sympathetic to us, than if they just go through abstract symbol manipulation to compute good or bad.

So, I think it's making it personal for the machines, for me, is an intriguing idea, making it experiencedly grounded within the machine, I think is an intriguing idea. And, again, part of that is motivated by learning, but I think it's a way of understanding, I think it's a fungible way of understanding others and I would like these machines to also have.

SI: Can a robot be conscious?

CB: I think a lot of it is that, even for ourselves, we don't really know what these terms mean and so, to be able to say is Machine X - you know, it forces us to really think hard about, "What do we really mean by that?"

And I think the closer we come to really being able to articulate what we really mean by that, means that it's that much more possible to start to try to develop machines that can really do that.

So, you can keep shuffling and moving the bar around. I think, for a case of these early robots, I'm willing, you know, like an insect for instance, I'm willing to grant it a sort of primordial, early hint of what eventually we would think of as being like, full-fledged consciousness. I can't believe it's a binary sort of thing, you either have it or you don't. I suspect it's a continuum.

I think part of the challenge before us is to try to understand what is that continuum and, again, really trying to hone our meaning of these terms. I think, certainly in a lot of conversations I've had with people, the one aspect that seems to be consistent is people, there's the sort of layperson kind of people, don't feel that it's possible to give a machine emotions.

I think what this basically really boils down to is the importance of the lifecycle. So, I might be able to program in these initial sort of emotional responses and networks and so forth but, fundamentally, that entity has to go through a lifecycle and have experiences that really ground out what those things are.

So, learning and growing, I think, for people, is a fundamental part of having genuine emotions, but I think that's certainly possible for a machine, that's certainly possible.

SI: How important are senses in creating machine "experience"?

Well, so, as far as experience, I think, for me for robot sensing, is the fundamental part of experience and I think this is when it gets interesting or tricky in that robots, as well, again, as other species, can have very different senses than what people have.

Which I think, again, kind of speaks to this, you know, we have to understand or kind of hone our thoughts on what emotion or consciousness are before we can start trying to map it to a dog's conscious or other specie's conscious, and whether a machine could be conscious.

And I think robots, particularly robots, are going to have a wide range of embodiments, a wide range of abilities and I think, certainly, one of the endeavors within robotics, and in many ways, it's not trying to replicate humans, because there's plenty of people.

We know how to make people. You want to build technologies that complement us. So, it's really, it is for me, much more about the partnership and synergy, rather than replicating. So, I think robots are, fundamentally, always going to be different from people; they're not human and they're never going to be human. And I think trying to measure them against the human yardstick is just kind of the wrong question.

So, again, for me, it's more about trying to understand what these things would mean when applied to a machine and trying to build machines in a way that can relate to us. That doesn't mean they have to be exactly like us, but it can at least relate to us.

SI: Can AI be created in a simulated world, or is sensory perception ncessary for intelligence?

I mean, first of all, I have to understand what people in that context would mean by an artificial general intelligence because you could probably develop a simulated world that's a very interesting world, that agents could acquire a wide variety of cognitive, what we would consider to be, intelligent behavior within that world.

But I would not go so far as that intelligence could then map to this world. And my work is fundamentally about this world. It's fundamentally about this physical world and people.

And so, for me, I think, certainly the nature of the embodiment and, for my case, for the importance of the physical robot. I think it's crucial to developing machines that understand this world and the full gamut of people. You know you don't really get the full breadth of human experience and behavior in simulated worlds. You know, we're just restricted by interfaces and all kinds of things.

So, again, I mean, I think the question is what do you mean by a general intelligence. I think if you're willing to grant that you can do a general intelligence artificially, I just don't think that's necessarily transfer to this world.

SI: What's exciting you in robotics?

CB:I think the real exciting thing that's happening within robotics, particularly social robotics, is that we are really starting to create machines that can start to relate to us and understand our behavior, not only in terms of our observable actions, but in terms of our thoughts and beliefs and intents.

And I think this core ability of social intelligence, of this beginnings of building machines that have a theory of other minds, I think is so fundamental to human social intelligence, so fundamental to how we can communicate and how we can learn from each other and how we can work with one another, that this is really opening up a whole new set of questions and abilities for what machines can do and how they can interact with us in the future.

And I think, if you extrapolate these kinds of questions and these kinds of social learnings and so forth, I think what this is pushing on, ultimately, is sort of the story of can machines think, to a more recent chapter, which I think we're still grappling with, which is can machines have emotion?

We're now starting to push on the sort of relational aspect and I think the next set of questions are going to be, when is a machine a person? When are we willing to grant the status of personhood to something that's not biologically human? And I think that's what, fundamentally, in the big, big picture, where this work is going to start challenging us in the future.